Merge remote-tracking branch 'upstream/develop' into merge-doc

.pre-commit-config.yaml

@@ -24,7 +24,7 @@
         description: Format files with ClangFormat.
         entry: clang-format -i
         language: system
-        files: \.(c|cc|cxx|cpp|h|hpp|hxx)$
+        files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx|proto)$
 -   repo: https://github.com/PaddlePaddle/pre-commit-golang
     sha: 8337620115c25ff8333f1b1a493bd031049bd7c0
     hooks:

CMakeLists.txt

@@ -36,8 +36,8 @@ include(simd)
 ################################ Configurations #######################################
 option(WITH_GPU         "Compile PaddlePaddle with NVIDIA GPU"          ${CUDA_FOUND})
 option(WITH_AVX         "Compile PaddlePaddle with AVX intrinsics"      ${AVX_FOUND})
-option(WITH_MKLDNN      "Compile PaddlePaddle with mkl-dnn support."    ${AVX_FOUND})
+option(WITH_MKLDNN      "Compile PaddlePaddle with mkl-dnn support."    OFF)
-option(WITH_MKLML       "Compile PaddlePaddle with mklml package."      ${AVX_FOUND})
+option(WITH_MKLML       "Compile PaddlePaddle with mklml package."      OFF)
 option(WITH_DSO         "Compile PaddlePaddle with dynamic linked CUDA" ON)
 option(WITH_TESTING     "Compile PaddlePaddle with unit testing"        ON)
 option(WITH_SWIG_PY     "Compile PaddlePaddle with inference api"       ON)

Dockerfile

@@ -27,13 +27,16 @@ RUN apt-get update && \
     git python-pip python-dev openssh-server bison  \
     wget unzip unrar tar xz-utils bzip2 gzip coreutils ntp \
     curl sed grep graphviz libjpeg-dev zlib1g-dev  \
-    python-numpy python-matplotlib gcc-4.8 g++-4.8 \
+    python-matplotlib gcc-4.8 g++-4.8 \
     automake locales clang-format-3.8 swig doxygen cmake  \
     liblapack-dev liblapacke-dev libboost-dev \
     clang-3.8 llvm-3.8 libclang-3.8-dev \
     net-tools && \
     apt-get clean -y
 
+# paddle is using numpy.flip, which is introduced since 1.12.0
+RUN pip --no-cache-dir install 'numpy>=1.12.0'
+
 # Install Go and glide
 RUN wget -O go.tgz https://storage.googleapis.com/golang/go1.8.1.linux-amd64.tar.gz && \
     tar -C /usr/local -xzf go.tgz && \

cmake/cpplint.cmake

@@ -56,11 +56,14 @@ macro(add_style_check_target TARGET_NAME)
                 # cpplint code style
                 get_filename_component(base_filename ${filename} NAME)
                 set(CUR_GEN ${CMAKE_CURRENT_BINARY_DIR}/${base_filename}.cpplint)
-                add_custom_command(TARGET ${TARGET_NAME} PRE_BUILD
+                add_custom_command(OUTPUT ${CUR_GEN} PRE_BUILD
                     COMMAND "${PYTHON_EXECUTABLE}" "${PROJ_ROOT}/paddle/scripts/cpplint.py"
                             "--filter=${STYLE_FILTER}"
                             "--write-success=${CUR_GEN}" ${filename}
+                    DEPENDS ${filename} ${PROJ_ROOT}/paddle/scripts/cpplint.py
                     WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR})
+                add_custom_target(${base_filename}.cpplint DEPENDS ${CUR_GEN})
+                add_dependencies(${TARGET_NAME} ${base_filename}.cpplint)
             endif()
         endforeach()
     endif()

cmake/external/gflags.cmake

@@ -28,7 +28,14 @@ INCLUDE_DIRECTORIES(${GFLAGS_INCLUDE_DIR})
 ExternalProject_Add(
     extern_gflags
     ${EXTERNAL_PROJECT_LOG_ARGS}
-    GIT_REPOSITORY  "https://github.com/gflags/gflags.git"
+    # TODO(yiwang): The annoying warnings mentioned in
+    # https://github.com/PaddlePaddle/Paddle/issues/3277 are caused by
+    # gflags.  I fired a PR https://github.com/gflags/gflags/pull/230
+    # to fix it.  Before it gets accepted by the gflags team, we use
+    # my personal fork, which contains above fix, temporarily.  Let's
+    # change this back to the official Github repo once my PR is
+    # merged.
+    GIT_REPOSITORY  "https://github.com/wangkuiyi/gflags.git"
     PREFIX          ${GFLAGS_SOURCES_DIR}
     UPDATE_COMMAND  ""
     CMAKE_ARGS      -DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}

cmake/external/python.cmake

@@ -24,7 +24,6 @@ IF(WITH_PYTHON)
 ENDIF(WITH_PYTHON)
 
 SET(py_env "")
-SET(USE_VIRTUALENV_FOR_TEST 1)
 IF(PYTHONINTERP_FOUND)
     find_python_module(pip REQUIRED)
     find_python_module(numpy REQUIRED)

cmake/generic.cmake

@@ -187,7 +187,13 @@ function(cc_library TARGET_NAME)
     endif()
     
     # cpplint code style
-    add_style_check_target(${TARGET_NAME} ${cc_library_SRCS})
+    foreach(source_file ${cc_library_SRCS})
+      string(REGEX REPLACE "\\.[^.]*$" "" source ${source_file})
+      if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h)
+        list(APPEND cc_library_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h)
+      endif()
+    endforeach()
+    add_style_check_target(${TARGET_NAME} ${cc_library_SRCS} ${cc_library_HEADERS})
 
   else(cc_library_SRCS)
     if (cc_library_DEPS)
@@ -239,6 +245,14 @@ function(nv_library TARGET_NAME)
         add_dependencies(${TARGET_NAME} ${nv_library_DEPS})
         target_link_libraries(${TARGET_NAME} ${nv_library_DEPS})
       endif()
+      # cpplint code style
+      foreach(source_file ${nv_library_SRCS})
+        string(REGEX REPLACE "\\.[^.]*$" "" source ${source_file})
+        if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h)
+          list(APPEND cc_library_HEADERS ${CMAKE_CURRENT_SOURCE_DIR}/${source}.h)
+        endif()
+      endforeach()
+      add_style_check_target(${TARGET_NAME} ${nv_library_SRCS} ${nv_library_HEADERS})
     else(nv_library_SRCS)
       if (nv_library_DEPS)
         merge_static_libs(${TARGET_NAME} ${nv_library_DEPS})

cmake/util.cmake

@@ -118,7 +118,6 @@ endfunction()
 macro(add_unittest_without_exec TARGET_NAME)
     add_executable(${TARGET_NAME} ${ARGN})
     link_paddle_test(${TARGET_NAME})
-    add_style_check_target(${TARGET_NAME} ${ARGN})
 endmacro()
 
 # add_unittest
@@ -150,9 +149,12 @@ endfunction()
 # Create a python unittest using run_python_tests.sh,
 # which takes care of making correct running environment
 function(add_python_test TEST_NAME)
-  add_test(NAME ${TEST_NAME}
+    foreach(arg ${ARGN})
-        COMMAND env PADDLE_PACKAGE_DIR=${PADDLE_PYTHON_PACKAGE_DIR}
+        get_filename_component(py_fn ${arg} NAME_WE)
-        bash ${PROJ_ROOT}/paddle/scripts/run_python_tests.sh
+        set(TRG_NAME ${TEST_NAME}_${py_fn})
-        ${USE_VIRTUALENV_FOR_TEST} ${PYTHON_EXECUTABLE} ${ARGN}
+        add_test(NAME ${TRG_NAME}
-        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR})
+                COMMAND env PYTHONPATH=${PADDLE_PYTHON_PACKAGE_DIR}
+                python2 ${arg}
+                WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR})
+    endforeach()
 endfunction()

paddle/.set_python_path.sh

@@ -21,22 +21,15 @@
 # 
 # It same as PYTHONPATH=${YOUR_PYTHON_PATH}:$PYTHONPATH {exec...}
 #
-
+PYPATH=""
-if ! python -c "import paddle" >/dev/null 2>/dev/null; then
+set -x
-  PYPATH=""
+while getopts "d:" opt; do
-  set -x
+  case $opt in
-  while getopts "d:" opt; do
+    d)
-    case $opt in
+      PYPATH=$OPTARG
-      d)
+      ;;
-        PYPATH=$OPTARG
+  esac
-        ;;
+done
-    esac
+shift $(($OPTIND - 1))
-  done
+export PYTHONPATH=$PYPATH:$PYTHONPATH
-  shift $(($OPTIND - 1))
+$@
-  export PYTHONPATH=$PYPATH:$PYTHONPATH
-  $@
-else
-  echo "paddle package is already in your PYTHONPATH. But unittest need a clean environment."
-  echo "Please uninstall paddle package before start unittest. Try to 'pip uninstall paddle'"
-  exit 1
-fi

paddle/CMakeLists.txt

@@ -15,7 +15,6 @@ if(Boost_FOUND)
   add_subdirectory(platform)
   add_subdirectory(framework)
   add_subdirectory(operators)
-  add_subdirectory(pybind)
 endif()
 
 if(WITH_C_API)

paddle/cuda/src/hl_batch_transpose.cu

@@ -12,17 +12,15 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include "hl_batch_transpose.h"
 #include "hl_base.h"
+#include "hl_batch_transpose.h"
 
 const int TILE_DIM = 64;
 const int BLOCK_ROWS = 16;
 
 // No bank-conflict transpose for a batch of data.
-__global__ void batchTransposeNoBankConflicts(real* odata,
+__global__ void batchTransposeNoBankConflicts(
-                                              const real* idata,
+    real* odata, const real* idata, int numSamples, int width, int height) {
-                                              int numSamples, int width,
-                                              int height) {
   __shared__ float tile[TILE_DIM][TILE_DIM + 1];
 
   const int x = blockIdx.x * TILE_DIM + threadIdx.x;
@@ -50,12 +48,12 @@ __global__ void batchTransposeNoBankConflicts(real* odata,
           newX] = tile[threadIdx.x][j];
 }
 
-void batchTranspose(const real* input, real* output, int width, int height,
+void batchTranspose(
-                    int batchSize) {
+    const real* input, real* output, int width, int height, int batchSize) {
   dim3 dimBlock(TILE_DIM, BLOCK_ROWS, 1);
   dim3 dimGrid(DIVUP(width, TILE_DIM), DIVUP(height, TILE_DIM), batchSize);
-  batchTransposeNoBankConflicts<<<dimGrid, dimBlock, 0, STREAM_DEFAULT>>>
+  batchTransposeNoBankConflicts<<<dimGrid, dimBlock, 0, STREAM_DEFAULT>>>(
-      (output, input, batchSize, width, height);
+      output, input, batchSize, width, height);
 
   CHECK_SYNC("batchTranspose failed!");
 }

paddle/cuda/src/hl_perturbation_util.cu

@@ -12,13 +12,12 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-
-#include <cmath>
 #include <stdlib.h>
-#include "hl_cuda.h"
+#include <cmath>
-#include "hl_time.h"
 #include "hl_base.h"
+#include "hl_cuda.h"
 #include "hl_perturbation_util.cuh"
+#include "hl_time.h"
 
 #define _USE_MATH_DEFINES
 
@@ -30,10 +29,16 @@ limitations under the License. */
  * centerX, centerY: translation.
  * sourceX, sourceY: output coordinates in the original image.
  */
-__device__ void getTranformCoord(int x, int y, real theta, real scale,
+__device__ void getTranformCoord(int x,
-                                 real tgtCenter, real imgCenter,
+                                 int y,
-                                 real centerR, real centerC,
+                                 real theta,
-                                 int* sourceX, int* sourceY) {
+                                 real scale,
+                                 real tgtCenter,
+                                 real imgCenter,
+                                 real centerR,
+                                 real centerC,
+                                 int* sourceX,
+                                 int* sourceY) {
   real H[4] = {cosf(-theta), -sinf(-theta), sinf(-theta), cosf(-theta)};
 
   // compute coornidates in the rotated and scaled image
@@ -57,11 +62,17 @@ __device__ void getTranformCoord(int x, int y, real theta, real scale,
  * created by Wei Xu (genome), converted by Jiang Wang
  */
 
-__global__ void kSamplingPatches(const real* imgs, real* targets,
+__global__ void kSamplingPatches(const real* imgs,
-                                 int imgSize, int tgtSize, const int channels,
+                                 real* targets,
-                                 int samplingRate, const real* thetas,
+                                 int imgSize,
-                                 const real* scales, const int* centerRs,
+                                 int tgtSize,
-                                 const int* centerCs, const real padValue,
+                                 const int channels,
+                                 int samplingRate,
+                                 const real* thetas,
+                                 const real* scales,
+                                 const int* centerRs,
+                                 const int* centerCs,
+                                 const real padValue,
                                  const int numImages) {
   const int caseIdx = blockIdx.x * 4 + threadIdx.x;
   const int pxIdx = blockIdx.y * 128 + threadIdx.y;
@@ -80,8 +91,15 @@ __global__ void kSamplingPatches(const real* imgs, real* targets,
     const int pxY = pxIdx / tgtSize;
 
     int srcPxX, srcPxY;
-    getTranformCoord(pxX, pxY, thetas[imgIdx], scales[imgIdx], tgtCenter,
+    getTranformCoord(pxX,
-                     imgCenter, centerCs[caseIdx], centerRs[caseIdx], &srcPxX,
+                     pxY,
+                     thetas[imgIdx],
+                     scales[imgIdx],
+                     tgtCenter,
+                     imgCenter,
+                     centerCs[caseIdx],
+                     centerRs[caseIdx],
+                     &srcPxX,
                      &srcPxY);
 
     imgs += (imgIdx * imgPixels + srcPxY * imgSize + srcPxX) * channels;
@@ -100,10 +118,15 @@ __global__ void kSamplingPatches(const real* imgs, real* targets,
  *
  * created by Wei Xu
  */
-void hl_generate_disturb_params(real*& gpuAngle, real*& gpuScaleRatio,
+void hl_generate_disturb_params(real*& gpuAngle,
-                                int*& gpuCenterR, int*& gpuCenterC,
+                                real*& gpuScaleRatio,
-                                int numImages, int imgSize, real rotateAngle,
+                                int*& gpuCenterR,
-                                real scaleRatio, int samplingRate,
+                                int*& gpuCenterC,
+                                int numImages,
+                                int imgSize,
+                                real rotateAngle,
+                                real scaleRatio,
+                                int samplingRate,
                                 bool isTrain) {
   // The number of output samples.
   int numPatches = numImages * samplingRate;
@@ -123,7 +146,8 @@ void hl_generate_disturb_params(real*& gpuAngle, real*& gpuScaleRatio,
     for (int i = 0; i < numImages; i++) {
       r_angle[i] =
           (rotateAngle * M_PI / 180.0) * (rand() / (RAND_MAX + 1.0)  // NOLINT
-                                          - 0.5);
+                                          -
+                                          0.5);
       s_ratio[i] =
           1 + (rand() / (RAND_MAX + 1.0) - 0.5) * scaleRatio;  // NOLINT
     }
@@ -140,8 +164,10 @@ void hl_generate_disturb_params(real*& gpuAngle, real*& gpuScaleRatio,
         int pxY =
             (int)(real(imgSize - 1) * rand() / (RAND_MAX + 1.0));  // NOLINT
 
-        const real H[4] = {cos(-r_angle[i]), -sin(-r_angle[i]),
+        const real H[4] = {cos(-r_angle[i]),
-                           sin(-r_angle[i]), cos(-r_angle[i])};
+                           -sin(-r_angle[i]),
+                           sin(-r_angle[i]),
+                           cos(-r_angle[i])};
         real x = pxX - imgCenter;
         real y = pxY - imgCenter;
         real xx = H[0] * x + H[1] * y;
@@ -185,9 +211,12 @@ void hl_generate_disturb_params(real*& gpuAngle, real*& gpuScaleRatio,
   delete[] center_c;
 }
 
-void hl_conv_random_disturb_with_params(const real* images, int imgSize,
+void hl_conv_random_disturb_with_params(const real* images,
-                                        int tgtSize, int channels,
+                                        int imgSize,
-                                        int numImages, int samplingRate,
+                                        int tgtSize,
+                                        int channels,
+                                        int numImages,
+                                        int samplingRate,
                                         const real* gpuRotationAngle,
                                         const real* gpuScaleRatio,
                                         const int* gpuCenterR,
@@ -202,29 +231,59 @@ void hl_conv_random_disturb_with_params(const real* images, int imgSize,
   dim3 threadsPerBlock(4, 128);
   dim3 numBlocks(DIVUP(numPatches, 4), DIVUP(targetSize, 128));
 
-  kSamplingPatches <<<numBlocks, threadsPerBlock>>>
+  kSamplingPatches<<<numBlocks, threadsPerBlock>>>(images,
-      (images, target, imgSize, tgtSize, channels, samplingRate,
+                                                   target,
-      gpuRotationAngle, gpuScaleRatio, gpuCenterR, gpuCenterC,
+                                                   imgSize,
-      paddingValue, numImages);
+                                                   tgtSize,
+                                                   channels,
+                                                   samplingRate,
+                                                   gpuRotationAngle,
+                                                   gpuScaleRatio,
+                                                   gpuCenterR,
+                                                   gpuCenterC,
+                                                   paddingValue,
+                                                   numImages);
 
   hl_device_synchronize();
 }
 
-void hl_conv_random_disturb(const real* images, int imgSize,
+void hl_conv_random_disturb(const real* images,
-                            int tgtSize, int channels, int numImages,
+                            int imgSize,
-                            real scaleRatio, real rotateAngle,
+                            int tgtSize,
-                            int samplingRate, real* gpu_r_angle,
+                            int channels,
-                            real* gpu_s_ratio, int* gpu_center_r,
+                            int numImages,
-                            int* gpu_center_c, int paddingValue,
+                            real scaleRatio,
-                            bool isTrain, real* targets) {
+                            real rotateAngle,
+                            int samplingRate,
+                            real* gpu_r_angle,
+                            real* gpu_s_ratio,
+                            int* gpu_center_r,
+                            int* gpu_center_c,
+                            int paddingValue,
+                            bool isTrain,
+                            real* targets) {
   // generate the random disturbance sequence and the sampling locations
-  hl_generate_disturb_params(gpu_r_angle, gpu_s_ratio, gpu_center_r,
+  hl_generate_disturb_params(gpu_r_angle,
-                  gpu_center_c, numImages, imgSize, rotateAngle,
+                             gpu_s_ratio,
-                  scaleRatio, samplingRate, isTrain);
+                             gpu_center_r,
-
+                             gpu_center_c,
-  hl_conv_random_disturb_with_params(
+                             numImages,
-                  images, imgSize, tgtSize, channels, numImages,
+                             imgSize,
-                  samplingRate, gpu_r_angle, gpu_s_ratio,
+                             rotateAngle,
-                  gpu_center_r, gpu_center_r, paddingValue,
+                             scaleRatio,
-                  targets);
+                             samplingRate,
+                             isTrain);
+
+  hl_conv_random_disturb_with_params(images,
+                                     imgSize,
+                                     tgtSize,
+                                     channels,
+                                     numImages,
+                                     samplingRate,
+                                     gpu_r_angle,
+                                     gpu_s_ratio,
+                                     gpu_center_r,
+                                     gpu_center_r,
+                                     paddingValue,
+                                     targets);
 }

paddle/cuda/src/hl_table_apply.cu

@@ -12,15 +12,16 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-
 #include "hl_base.h"
-#include "hl_device_functions.cuh"
 #include "hl_cuda.h"
+#include "hl_device_functions.cuh"
 #include "paddle/utils/Logging.h"
 
-template<int blockDimX, int blockDimY, int gridDimX, bool AddRow>
+template <int blockDimX, int blockDimY, int gridDimX, bool AddRow>
-__global__ void KeMatrixAddRows(real* output, int ldo,
+__global__ void KeMatrixAddRows(real* output,
-                                real* table, int ldt,
+                                int ldo,
+                                real* table,
+                                int ldt,
                                 int* ids,
                                 int numSamples,
                                 int tableSize,
@@ -31,8 +32,8 @@ __global__ void KeMatrixAddRows(real* output, int ldo,
   while (idy < numSamples) {
     int tableId = ids[idy];
     if ((0 <= tableId) && (tableId < tableSize)) {
-      real *out = output + idy * ldo;
+      real* out = output + idy * ldo;
-      real *tab = table + tableId * ldt;
+      real* tab = table + tableId * ldt;
       for (int i = idx; i < dim; i += blockDimX) {
         if (AddRow) {
           paddle::paddleAtomicAdd(&tab[i], out[i]);
@@ -45,8 +46,10 @@ __global__ void KeMatrixAddRows(real* output, int ldo,
   }
 }
 
-void hl_matrix_select_rows(real* output, int ldo,
+void hl_matrix_select_rows(real* output,
-                           real* table, int ldt,
+                           int ldo,
+                           real* table,
+                           int ldt,
                            int* ids,
                            int numSamples,
                            int tableSize,
@@ -57,14 +60,16 @@ void hl_matrix_select_rows(real* output, int ldo,
 
   dim3 threads(128, 8);
   dim3 grid(8, 1);
-  KeMatrixAddRows<128, 8, 8, 0><<< grid, threads, 0, STREAM_DEFAULT >>>
+  KeMatrixAddRows<128, 8, 8, 0><<<grid, threads, 0, STREAM_DEFAULT>>>(
-    (output, ldo, table, ldt, ids, numSamples, tableSize, dim);
+      output, ldo, table, ldt, ids, numSamples, tableSize, dim);
 
   CHECK_SYNC("hl_matrix_select_rows failed");
 }
 
-void hl_matrix_add_to_rows(real* table, int ldt,
+void hl_matrix_add_to_rows(real* table,
-                           real* input, int ldi,
+                           int ldt,
+                           real* input,
+                           int ldi,
                            int* ids,
                            int numSamples,
                            int tableSize,
@@ -75,16 +80,15 @@ void hl_matrix_add_to_rows(real* table, int ldt,
 
   dim3 threads(128, 8);
   dim3 grid(8, 1);
-  KeMatrixAddRows<128, 8, 8, 1><<< grid, threads, 0, STREAM_DEFAULT >>>
+  KeMatrixAddRows<128, 8, 8, 1><<<grid, threads, 0, STREAM_DEFAULT>>>(
-    (input, ldi, table, ldt, ids, numSamples, tableSize, dim);
+      input, ldi, table, ldt, ids, numSamples, tableSize, dim);
 
   CHECK_SYNC("hl_matrix_add_to_rows failed");
 }
 
-template<class T, int blockDimX, int gridDimX>
+template <class T, int blockDimX, int gridDimX>
-__global__ void KeVectorSelect(T* dst, int sized,
+__global__ void KeVectorSelect(
-                               const T* src, int sizes,
+    T* dst, int sized, const T* src, int sizes, const int* ids, int sizei) {
-                               const int* ids, int sizei) {
   int idx = threadIdx.x + blockDimX * blockIdx.x;
   while (idx < sizei) {
     int index = ids[idx];
@@ -95,9 +99,8 @@ __global__ void KeVectorSelect(T* dst, int sized,
 }
 
 template <class T>
-void hl_vector_select_from(T* dst, int sized,
+void hl_vector_select_from(
-                           const T* src, int sizes,
+    T* dst, int sized, const T* src, int sizes, const int* ids, int sizei) {
-                           const int* ids, int sizei) {
   CHECK_NOTNULL(dst);
   CHECK_NOTNULL(src);
   CHECK_NOTNULL(ids);
@@ -105,18 +108,17 @@ void hl_vector_select_from(T* dst, int sized,
 
   dim3 threads(512, 1);
   dim3 grid(8, 1);
-  KeVectorSelect<T, 512, 8><<< grid, threads, 0, STREAM_DEFAULT >>>
+  KeVectorSelect<T, 512, 8><<<grid, threads, 0, STREAM_DEFAULT>>>(
-    (dst, sized, src, sizes, ids, sizei);
+      dst, sized, src, sizes, ids, sizei);
 
   CHECK_SYNC("hl_vector_select_from failed");
 }
 
-template
+template void hl_vector_select_from(real* dst,
-void hl_vector_select_from(real* dst, int sized,
+                                    int sized,
-                           const real* src, int sizes,
+                                    const real* src,
-                           const int* ids, int sizei);
+                                    int sizes,
-template
+                                    const int* ids,
-void hl_vector_select_from(int* dst, int sized,
+                                    int sizei);
-                           const int* src, int sizes,
+template void hl_vector_select_from(
-                           const int* ids, int sizei);
+    int* dst, int sized, const int* src, int sizes, const int* ids, int sizei);
-

paddle/framework/CMakeLists.txt

@@ -12,13 +12,15 @@ cc_test(variable_test SRCS variable_test.cc)
 cc_library(scope SRCS scope.cc)
 cc_test(scope_test SRCS scope_test.cc DEPS scope)
 
-proto_library(attr_type SRCS attr_type.proto)
+proto_library(attribute_proto SRCS attribute.proto)
-proto_library(op_proto SRCS op_proto.proto DEPS attr_type)
+proto_library(op_proto SRCS op_proto.proto DEPS attribute_proto)
-proto_library(op_desc SRCS op_desc.proto DEPS attr_type)
+proto_library(op_desc SRCS op_desc.proto DEPS attribute_proto)
 cc_test(op_proto_test SRCS op_proto_test.cc DEPS op_proto protobuf)
 cc_test(op_desc_test SRCS op_desc_test.cc DEPS op_desc protobuf)
 
-cc_library(operator SRCS operator.cc DEPS op_desc device_context tensor scope)
+cc_library(attribute SRCS attribute.cc DEPS op_desc op_proto)
+
+cc_library(operator SRCS operator.cc DEPS op_desc device_context tensor scope attribute)
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry)
 
 cc_library(grad_op_builder SRCS grad_op_builder.cc DEPS op_proto operator)
@@ -26,13 +28,19 @@ cc_library(op_registry SRCS op_registry.cc DEPS op_desc grad_op_builder)
 cc_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
 cc_test(grad_op_builder_test SRCS grad_op_builder_test.cc DEPS grad_op_builder op_registry add_op)
 
-py_proto_compile(framework_py_proto SRCS attr_type.proto op_proto.proto op_desc.proto)
+py_proto_compile(framework_py_proto SRCS attribute.proto op_proto.proto op_desc.proto)
 # Generate an empty __init__.py to make framework_py_proto as a valid python module.
 add_custom_target(framework_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py)
 add_dependencies(framework_py_proto framework_py_proto_init)
 
-cc_library(net SRCS net.cc DEPS op_registry)
+cc_library(backward SRCS backward.cc DEPS net_op)
-cc_test(net_op_test SRCS net_op_test.cc DEPS net)
-
-cc_library(backward SRCS backward.cc DEPS net)
 cc_test(backward_test SRCS backward_test.cc DEPS backward)
+cc_library(paddle_pybind SHARED
+    SRCS pybind.cc
+    DEPS pybind python backward
+	fc_op
+	sgd_op
+	add_op
+	mean_op
+	cross_entropy_op
+	recurrent_op)

paddle/framework/attribute.cc

@@ -0,0 +1,85 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/framework/attribute.h"
+
+#include <vector>
+
+namespace paddle {
+namespace framework {
+
+template <>
+AttrType AttrTypeID<int>() {
+  return INT;
+}
+template <>
+AttrType AttrTypeID<float>() {
+  return FLOAT;
+}
+template <>
+AttrType AttrTypeID<std::string>() {
+  return STRING;
+}
+template <>
+AttrType AttrTypeID<std::vector<int>>() {
+  return INTS;
+}
+template <>
+AttrType AttrTypeID<std::vector<float>>() {
+  return FLOATS;
+}
+template <>
+AttrType AttrTypeID<std::vector<std::string>>() {
+  return STRINGS;
+}
+
+Attribute GetAttrValue(const AttrDesc& attr_desc) {
+  switch (attr_desc.type()) {
+    case paddle::framework::AttrType::INT: {
+      return attr_desc.i();
+    }
+    case paddle::framework::AttrType::FLOAT: {
+      return attr_desc.f();
+    }
+    case paddle::framework::AttrType::STRING: {
+      return attr_desc.s();
+    }
+    case paddle::framework::AttrType::INTS: {
+      std::vector<int> val(attr_desc.ints_size());
+      for (int i = 0; i < attr_desc.ints_size(); ++i) {
+        val[i] = attr_desc.ints(i);
+      }
+      return val;
+    }
+    case paddle::framework::AttrType::FLOATS: {
+      std::vector<float> val(attr_desc.floats_size());
+      for (int i = 0; i < attr_desc.floats_size(); ++i) {
+        val[i] = attr_desc.floats(i);
+      }
+      return val;
+    }
+    case paddle::framework::AttrType::STRINGS: {
+      std::vector<std::string> val(attr_desc.strings_size());
+      for (int i = 0; i < attr_desc.strings_size(); ++i) {
+        val[i] = attr_desc.strings(i);
+      }
+      return val;
+    }
+  }
+  PADDLE_ENFORCE(false, "Unknown OpDesc::AttrDesc::type !");
+  return boost::blank();
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/attribute.h

@@ -1,3 +1,17 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
 #pragma once
 
 #include <boost/variant.hpp>
@@ -6,6 +20,9 @@
 #include <unordered_map>
 #include <unordered_set>
 #include <vector>
+
+#include "paddle/framework/attribute.pb.h"
+#include "paddle/framework/op_desc.pb.h"
 #include "paddle/platform/enforce.h"
 
 namespace paddle {
@@ -14,13 +31,19 @@ namespace framework {
 typedef boost::variant<boost::blank, int, float, std::string, std::vector<int>,
                        std::vector<float>, std::vector<std::string>>
     Attribute;
+
 typedef std::unordered_map<std::string, Attribute> AttributeMap;
 
+template <typename T>
+AttrType AttrTypeID();
+
+Attribute GetAttrValue(const AttrDesc& attr_desc);
+
 // check whether a value(attribute) fit a certain limit
 template <typename T>
 class LargerThanChecker {
  public:
-  LargerThanChecker(T lower_bound) : lower_bound_(lower_bound) {}
+  explicit LargerThanChecker(T lower_bound) : lower_bound_(lower_bound) {}
   void operator()(T& value) const {
     PADDLE_ENFORCE(value > lower_bound_, "larger_than check fail");
   }
@@ -35,7 +58,8 @@ class LargerThanChecker {
 template <typename T>
 class DefaultValueSetter {
  public:
-  DefaultValueSetter(T default_value) : default_value_(default_value) {}
+  explicit DefaultValueSetter(T default_value)
+      : default_value_(default_value) {}
   void operator()(T& value) const { value = default_value_; }
 
  private:
@@ -78,7 +102,8 @@ class TypedAttrChecker {
   typedef std::function<void(T&)> ValueChecker;
 
  public:
-  TypedAttrChecker(const std::string& attr_name) : attr_name_(attr_name) {}
+  explicit TypedAttrChecker(const std::string& attr_name)
+      : attr_name_(attr_name) {}
 
   TypedAttrChecker& InEnum(const std::unordered_set<T>& range) {
     value_checkers_.push_back(EnumInContainer<T>(range));

paddle/framework/attribute.proto

@@ -12,17 +12,17 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-syntax="proto2";
+syntax = "proto2";
 package paddle.framework;
 
 // Attribute Type for paddle's Op.
 // Op contains many attributes. Each type of attributes could be different.
 // The AttrType will be shared between AttrDesc and AttrProto.
 enum AttrType {
-    INT = 0;
+  INT = 0;
-    FLOAT = 1;
+  FLOAT = 1;
-    STRING = 2;
+  STRING = 2;
-    INTS = 3;
+  INTS = 3;
-    FLOATS = 4;
+  FLOATS = 4;
-    STRINGS = 5;
+  STRINGS = 5;
 }

paddle/framework/backward.cc

@@ -14,8 +14,8 @@
 
 #include "paddle/framework/backward.h"
 #include <list>
-#include "paddle/framework/net.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/operators/net_op.h"
 
 namespace paddle {
 namespace framework {
@@ -32,7 +32,7 @@ static bool AllInSet(const std::vector<std::string>& names,
 }
 
 static std::shared_ptr<OperatorBase> NOP() {
-  auto net_op = std::make_shared<NetOp>();
+  auto net_op = std::make_shared<operators::NetOp>();
   net_op->type_ = "@NOP@";
   net_op->CompleteAddOp();
   return net_op;
@@ -42,9 +42,9 @@ static std::shared_ptr<OperatorBase> NOP() {
 //
 //  no_grad_names the gradient variable names without gradient calculating.
 //
-//  uniq_id is a unique index used inside recursively calling BackwardRecursive.
+//  uniq_id is a unique index used inside recursively calling
-//  use `uid = uniq_id++;` to get the unique index, and pass `uniq_id` through
+//  BackwardRecursive. use `uid = uniq_id++;` to get the unique index, and
-//  recursive calling.
+//  pass `uniq_id` through recursive calling.
 //
 //  returns The backward operator. For simple situation, it is a simple
 //  operator. For complex situation, it is a NetOp.
@@ -59,32 +59,30 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
   //  If all input gradients of forwarding operator do not need to calculate,
   //  just return an NOP. Not return null ptr because NOP does not take
   //  too much time for calculation, but it is useful for simplifying logic.
-  if (AllInSet(forwardOp.inputs_, OperatorBase::GRAD_VAR_SUFFIX(),
+  if (AllInSet(forwardOp.inputs_, kGradVarSuffix, no_grad_names)) {
-               no_grad_names)) {
     return NOP();
   }
 
-  //  All output gradients of forwarding operator do not need to calculate. Then
+  //  All output gradients of forwarding operator do not need to calculate.
-  //  all input gradients cannot be computed at all, and we put them into
+  //  Then all input gradients cannot be computed at all, and we put them into
   //  `no_grad_names` set. Return an NOP.
-  if (AllInSet(forwardOp.outputs_, OperatorBase::GRAD_VAR_SUFFIX(),
+  if (AllInSet(forwardOp.outputs_, kGradVarSuffix, no_grad_names)) {
-               no_grad_names)) {
     for (auto& name : forwardOp.inputs_) {
       // Mark all input is not need
-      no_grad_names.insert(name + OperatorBase::GRAD_VAR_SUFFIX());
+      no_grad_names.insert(name + kGradVarSuffix);
     }
     return NOP();
   }
 
   // Returned gradient network
-  auto net = std::make_shared<NetOp>();
+  auto net = std::make_shared<operators::NetOp>();
 
   if (forwardOp.IsNetOp()) {
     // Because forwardOp is a net op, it can static_cast.
-    auto& forwardNet = static_cast<const NetOp&>(forwardOp);
+    auto& forwardNet = static_cast<const operators::NetOp&>(forwardOp);
 
-    // Map from output gradient variable name to operator's indices in backward
+    // Map from output gradient variable name to operator's indices in
-    // net. That operator generates that variable.
+    // backward net. That operator generates that variable.
     std::unordered_map<std::string, std::vector<size_t>> dup_output_ops;
 
     size_t local_op_id = 0;
@@ -134,9 +132,9 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
     std::shared_ptr<OperatorBase> grad_op = OpRegistry::CreateGradOp(forwardOp);
     for (std::string& grad_input : grad_op->inputs_) {
       if (no_grad_names.count(grad_input)) {
-        std::string prefix = grad_input.substr(
+        std::string prefix =
-            0, grad_input.size() - OperatorBase::GRAD_VAR_SUFFIX().size());
+            grad_input.substr(0, grad_input.size() - kGradVarSuffix.size());
-        grad_input = prefix + OperatorBase::ZERO_VAR_SUFFIX();
+        grad_input = prefix + kZeroVarSuffix;
 
         // If part of input gradient of that operator is not calculated, fill
         // zero variables to that input gradient.
@@ -147,7 +145,7 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
 
     for (std::string& grad_output : grad_op->outputs_) {
       if (no_grad_names.count(grad_output)) {
-        grad_output = OperatorBase::EMPTY_VAR_NAME();
+        grad_output = kEmptyVarName;
       }
     }
 
@@ -168,14 +166,14 @@ std::shared_ptr<OperatorBase> Backward(
   std::unordered_set<std::string> no_grad_names;
   no_grad_names.reserve(no_grad_vars.size());
 
-  no_grad_names.insert(OperatorBase::EMPTY_VAR_NAME() +
+  no_grad_names.insert(kEmptyVarName + kGradVarSuffix);
-                       OperatorBase::GRAD_VAR_SUFFIX());
 
   for (auto& name : no_grad_vars) {
-    no_grad_names.insert(name + OperatorBase::GRAD_VAR_SUFFIX());
+    no_grad_names.insert(name + kGradVarSuffix);
   }
   size_t uid = 0;
   return BackwardRecursive(forwardOp, no_grad_names, uid);
 }
+
 }  // namespace framework
 }  // namespace paddle